From 8c11ddfc5bf9770bbc46ae2168cf132a50340e46 Mon Sep 17 00:00:00 2001
From: Yuta Saito <uc4w6c@bma.biglobe.ne.jp>
Date: Fri, 9 Jan 2026 17:12:17 +0900
Subject: [PATCH] fix: periodically refresh MCP registry across instances via
 scheduler job

---
 .../mcp_server/mcp_server_manager.py          | 28 +++++++++++++++---
 litellm/proxy/proxy_server.py                 | 29 +++++++++++++++++++
 .../types/mcp_server/mcp_server_manager.py    |  2 ++
 3 files changed, 55 insertions(+), 4 deletions(-)

diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
index 3a548e203c5..6e7db26a95c 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
@@ -551,6 +551,7 @@ class MCPServerManager:
             allowed_tools=getattr(mcp_server, "allowed_tools", None),
             disallowed_tools=getattr(mcp_server, "disallowed_tools", None),
             allow_all_keys=mcp_server.allow_all_keys,
+            updated_at=getattr(mcp_server, "updated_at", None),
         )
         return new_server
 
@@ -2074,15 +2075,34 @@ class MCPServerManager:
         db_mcp_servers = await get_all_mcp_servers(prisma_client)
         verbose_logger.info(f"Found {len(db_mcp_servers)} MCP servers in database")
 
-        # ensure the global_mcp_server_manager is up to date with the db
+        previous_registry = self.registry
+        new_registry: Dict[str, MCPServer] = {}
+
         for server in db_mcp_servers:
+            existing_server = previous_registry.get(server.server_id)
+
+            if (
+                existing_server is not None
+                and existing_server.updated_at is not None
+                and server.updated_at is not None
+                and existing_server.updated_at == server.updated_at
+            ):
+                # Re-use existing server instance to avoid re-running build_mcp_server_from_table()
+                # which can perform network discovery for OAuth2 servers.
+                new_registry[server.server_id] = existing_server
+                continue
+
             verbose_logger.debug(
-                f"Adding server to registry: {server.server_id} ({server.server_name})"
+                f"Building server from DB: {server.server_id} ({server.server_name})"
             )
-            await self.add_server(server)
+            new_registry[server.server_id] = await self.build_mcp_server_from_table(
+                server
+            )
+
+        self.registry = new_registry
 
         verbose_logger.debug(
-            f"Registry now contains {len(self.get_registry())} servers"
+            "MCP registry refreshed (%s servers in registry)", len(new_registry)
         )
 
     def get_mcp_servers_from_ids(self, server_ids: List[str]) -> List[MCPServer]:
diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index 0e58ae7c643..77a0f5a1c09 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -4109,6 +4109,23 @@ class ProxyConfig:
             return []
 
 
+async def _reload_mcp_servers_job():
+    """Background job entrypoint for MCP registry refreshes."""
+    if proxy_config._should_load_db_object(object_type="mcp") is False:
+        return
+
+    try:
+        from litellm.proxy._experimental.mcp_server.mcp_server_manager import (
+            global_mcp_server_manager,
+        )
+
+        await global_mcp_server_manager._add_mcp_servers_from_db_to_in_memory_registry()  # noqa: SLF001
+    except Exception as e:
+        verbose_proxy_logger.exception(
+            "Failed to reload MCP servers from database: %s", str(e)
+        )
+
+
 proxy_config = ProxyConfig()
 
 
@@ -4646,6 +4663,18 @@ class ProxyStartupEvent:
                 misfire_grace_time=APSCHEDULER_MISFIRE_GRACE_TIME,
             )
             await proxy_config.get_credentials(prisma_client=prisma_client)
+
+            from litellm.proxy._experimental.mcp_server.utils import is_mcp_available
+
+            if is_mcp_available():
+                scheduler.add_job(
+                    _reload_mcp_servers_job,
+                    "interval",
+                    seconds=30,
+                    id="reload_mcp_servers_job",
+                    replace_existing=True,
+                    misfire_grace_time=APSCHEDULER_MISFIRE_GRACE_TIME,
+                )
         await cls._initialize_slack_alerting_jobs(
             scheduler=scheduler,
             general_settings=general_settings,
diff --git a/litellm/types/mcp_server/mcp_server_manager.py b/litellm/types/mcp_server/mcp_server_manager.py
index 96fd79f466b..94f33ffb297 100644
--- a/litellm/types/mcp_server/mcp_server_manager.py
+++ b/litellm/types/mcp_server/mcp_server_manager.py
@@ -1,3 +1,4 @@
+from datetime import datetime
 from typing import Any, Dict, List, Optional
 
 from pydantic import BaseModel, ConfigDict
@@ -50,4 +51,5 @@ class MCPServer(BaseModel):
     env: Optional[Dict[str, str]] = None
     access_groups: Optional[List[str]] = None
     allow_all_keys: bool = False
+    updated_at: Optional[datetime] = None
     model_config = ConfigDict(arbitrary_types_allowed=True)